SECOND-ORDER LOGIC:
ONTOLOGICAL AND EPISTEMOLOGICAL PROBLEMS
Marcus Rossberg
A Thesis Submitted for the Degree of PhD at the
University of St Andrews
2006
Full metadata for this item is available in Research@StAndrews:FullText
at:
http://research-repository.st-andrews.ac.uk/
Please use this identifier to cite or link to this item:
Second-Order Logic:
Ontological and Epistemological Problems
Marcus Rossberg
Supervisor: Professor Crispin Wright
Second supervisor: Professor Stewart Shapiro
Date of submission: January 5th, 2006
Date of final examination: March 4th, 2006
Examiners: Dr Peter Clark
F¨ur meine Eltern, Irene und Wolfgang Rossberg,
meine Großm¨utter, Hedwig Grosche und Ilse Rossberg,
und gewidmet dem Andenken an meinen Großvater
Paul Rossberg
(1920 – 2004)
For my parents, Irene and Wolfgang Rossberg,
my grandmothers, Hedwig Grosche and Ilse Rossberg,
and dedictated to the memory of my grandfather
Paul Rossberg
Abstract
In this thesis I provide a survey over different approaches to second-order logic
and its interpretation, and introduce a novel approach. Of special interest are the
questions whether (a particular form of) second-order logic can count as logic in some
(further to be specified) proper sense of logic, and what epistemic status it occupies.
More specifically, second-order logic is sometimes taken to be mathematical, a mere
notational variant of some fragment of set theory. If this is the case, it might be
argued that it does not have the “epistemic innocence” which would be needed
for, e.g., foundational programmes in (the philosophy of) mathematics for which
second-order logic is sometimes used. I suggest a Deductivist conception of logic,
that characterises logical consequence by means of inference rules, and argue that
Acknowledgements
My primary thanks go to my supervisors. Crispin Wright provided comments, crit-icism, discussion, support, and more criticism whenever I wanted or needed it. His dedication to my supervision, in particular in the final months, was beyond the call of duty. I also greatly benefitted from the comments and criticism I got from Stew-art Shapiro during his annual visits to St Andrews and when I went to visit him in Columbus in Spring 2004. Both my supervisors let me run with my own ideas when I needed the space to develop them, and demanded rigour and “tidying up” once I had run far enough. I am also indebted to my shadow supervisors Fraser MacBride and Roy Cook. The feedback I got from them, in particular in the the first couple of years, was invaluable.
I was lucky enough to be able to write this thesis as a member of the Arch´e research centre. The active and collaborative atmosphere here creates what is prob-ably a unique environment for philosophical research. First and foremost amongst my fellows here I want to thank Philip “Kollege” Ebert and Nikolaj Jang Peder-sen who together with me were the first generation ofArch´e postgraduate students. We were soon joined by the next two “Arch´e Beavers”, Ross Cameron and Robbie Williams. One cannot hope for better fellow Ph.D. students. Thanks guys! The thanks extends to the other members of Arch´e: Sama Agahi, Bob Hale, Paul Mc-Callion, Darren McDonald, Sean Morris, Agust´ın Rayo, Andrea Sereni, and Chiara Tabet in the Neo-Fregean project, and my other colleagues in the centre, Elizabeth Barnes, Eline Busck, Richard Dietz, Patrick Greenough, Aviv Hoffmann, Carrie Jenkins, Dan L´opez de Sa, Sebastiano Moruzzi, Josh Parsons, Graham Priest, Anna Sherratt, S´onia Roca, and Elia Zardini.
The friendly, collaborative, but also philosophically intense atmosphere was not restricted toArch´e, however; it is a also a characteristic mark of the whole postgrad-uate community in St Andrews. The Friday Seminar, which we all lived through together, provided an excellent training ground in all respects, philosophical and social (the latter in particular following the seminar, in the pub).
during the very stressful last few weeks before the submission of this thesis – weeks that did not seem to end.
Much of the material in this thesis was first tested in the Friday seminar as well as the Arch´e research seminar. I am grateful for the many comments and criti-cisms I received. Earlier versions of some of the chapters were presented at various conferences, workshops and seminars all over Europe, including the Universities of D¨usseldorf, Helsinki, and Stockholm, the Ockham Society in Oxford, the SPPA con-ference in Stirling, the GAP.5 in Bielefeld, the LMPS03 in Oviedo, the FOL’75 in Berlin, and the ECAP5 in Lisbon. Many thanks to the audiences there. Chapter 5 profited greatly from a discussion with Kit Fine to whom I am indebted for his comments and suggestions.
I am grateful for the financial support I received: for my tuition fees from the Arts and Humanties Research Board (AHRB – now AHRC), a full maintenance scholarship from the Studienstiftung des Deutschen Volkes [German National Aca-demic Foundation] from my second year on, and a travel award from the Russell Trustfor a visiting scholarship at the Ohio State University at Columbus.
I could not have written this thesis without the support of my family. There are no words for my gratitude towards my parents, Irene and Wolfgang Rossberg, and my grandparents, Hedwig Grosche and Ilse and Paul Rossberg. My grandfather Paul Rossberg sadly did not live to see the completion of my postgraduate studies.
Notational Conventions
For the benefit of a consistent use of the logical symbols I have tacitly changed
the symbols other authors use into the ones preferred here. Boolos’ ‘→’ and ‘↔’,
for example, are replaced by ‘⊃’ and ‘≡’, respectively. Likewise, Quine’s ‘(x)’, for
example, is replaced by ‘∀x’, and his notation using ‘.’, ‘:’, ‘.:’, etc., as both symbols
for conjunction and for scope distinctions is abandoned in favour of the nowadays
more common use of ‘∧’ (for conjunction) and the use of parentheses (for scope
distincitons).
I use single quotations marks for mentioning an expression, and double
quota-tions marks where the quoted expression is used; the latter case almost exclusively
occurs for verbatim quotations from the literature and the use of a word in a
non-literal sense or in a sense that I explicitly do not agree with. It should in all cases be
clear from the context in which of these senses I uses the double quotation marks. I
tacitly changed the quotations marks in verbatim quotations from the literature to
accord to this rule where authors followed different conventions. (Quine and Boolos,
for example, sometimes use double quotation marks for mentioning expressions.)
Where longer passages are quoted verbatim from the literature, the quotation
is separated from the main text by an extra wide margin, and quotation marks are
omitted.
Corner quotes, ‘p’ and ‘q’, are used as devices for quasi-quotation, as introduced
Contents
1 Introduction 1
1.1 The Project . . . 1
1.2 Outline of the Thesis . . . 5
2 The Formal System of Second-Order Logic 9 2.1 Preliminaries . . . 9
2.2 Language . . . 9
2.3 Deductive Systems . . . 11
2.3.1 Axiomatic System . . . 11
2.3.2 Natural Deduction . . . 14
2.4 Semantics . . . 16
2.4.1 Semantics for First-Order Logic . . . 16
2.4.2 Standard Semantics for Second-Order Logic . . . 17
2.4.3 Henkin Semantics . . . 18
2.5 Some Meta-Theoretic Results . . . 19
2.5.1 Standard and Henkin Semantics . . . 20
2.5.2 First-Order Logic . . . 20
2.5.3 Second-Order Logic with Standard Semantics . . . 22
2.5.4 Second-Order Logic with Henkin Semantics . . . 24
3 On Quine 26 3.1 Introduction . . . 26
3.2 Incompleteness and Branching Quantifiers . . . 29
3.3 Incoherence and Unintelligibility . . . 35
3.3.1 Russell’s Paradox . . . 35
3.3.2 No Entity Without Identity . . . 39
3.4 Dishonesty: The Hidden Staggering Ontology . . . 42
4 Plurals 52 4.1 Some Critics . . . 52
4.3 Polyadic Predicates . . . 68
4.4 Still Wild . . . 75
4.5 Summary . . . 86
5 Ontological Commitment 89 5.1 Quine’s Criterion . . . 89
5.2 Universals . . . 94
5.3 Second-Order Quantification . . . 95
5.4 Polyadic Predicates Again . . . 100
5.5 A New Criterion . . . 105
6 Semantic Incompleteness∗ 112 6.1 Introduction . . . 112
6.2 Logical Consequence . . . 114
6.3 Refining the Picture . . . 120
6.4 Some Arguments for Completeness . . . 127
6.5 Conclusion . . . 143
7 The Semantic Conception 145 7.1 Introduction . . . 145
7.2 Mathematical Theories and their Intended Interpretations . . . 148
7.2.1 Categoricity . . . 149
7.2.2 Embedding . . . 154
7.3 Expressive Power . . . 157
7.3.1 Cardinalities . . . 158
7.3.2 The Continuum Hypothesis . . . 163
7.4 Substantial Content . . . 167
7.5 Conclusion . . . 180
8 The Deductivist Conception 182 8.1 Introduction . . . 182
8.2 Themes from Frege . . . 183
8.3 Purely Syntactic Approaches . . . 193
8.4 The Deductivist Approach . . . 197
8.5 Two Problems for the Deductivist Approach . . . 208
8.5.1 Inherent Incompleteness . . . 208
8.5.2 Impredicativity . . . 222
9.2 Etchemendy’s Strategy . . . 234
9.3 Conceptual Inadequacy . . . 237
9.4 Extensional Inadequacy . . . 245
9.5 Representational Semantics . . . 248
9.5.1 Historical Intermission . . . 255
9.6 Comparisons . . . 257
Chapter 1
Introduction
1.1
The Project
The argument of this thesis is for the claim that second-order logic isproper logic.
This is not merely a question of handing out honorifics, more or less arbitrarily,
but a substantial claim about the nature of second-order logic, properly construed.
Proper logic plays a crucial role in codifying correct inference. While there are, of
course, various species of correct inferences, including mathematical ways of
reason-ing, the epistemic advantage of a proper logic is that the conclusion of an argument,
that is valid in the proper logical way, is ideally justified on basis of its premises.
This concept was introduced by Stephen Wagner in his (Wagner, 1987). Ideal
jus-tification is Wagner’s way of spelling out a thought contained already in Gottlob
Frege’s writings that logic is not supposed to add anything to the inference.
Ev-erything that is needed to justify a claim is already contained in the premises; only
then is an inference properly logical.
and thus cannot count as a proper logic. I will argue against this claim, and also
against other allegations against second-order logic that have been put forward.
Most commonly it is believed that second-order logic is really a mathematical theory,
very much akin to set theory. This claim goes back to W.V. Quine’s quip that
second-order logic is “set theory in sheep’s clothing”.1 The problem with this is
not that set theory is false, but that it is a strong mathematical theory, and not
logic. Set theory makes enough substantial mathematical claims that virtually all
areas of mathematics can be represented in it. Ideal justification is not in general
possible with such a system, since these substantial mathematical claims go into the
argument from premises to conclusion, and we cannot in general be sure that the
conclusion is indeed solely justified in basis of the premises, or if it partly rests on
some of these substantial mathematical presupposition.
Since mathematical truths are, presumably, necessarily true, if true at all, why
would that matter? It does not always matter. Mathematics is applicable in the
sciences, for example, and presumably does not cause any problems there. Indeed,
many think it is indispensable. There are areas, on the other hand, where no
mathematics should be presupposed. The philosophy of mathematics is one such
area, at least as construed by some research projects in this area. Hartry Field’s
nominalist programme proposed in his Science Without Numbers2, for example,
argues against the claim that mathematics is indispensable to science. His case
study is Newtonian Mechanics, for which he produces a formal system that does
not contain any mathematics. Formal systems are built on a formal logic, however,
and if the alleged logic that is used would itself really be a mathematical theory, the
argument would break down. Field indeed considers the advantages a second-order
formulation of his system would have, but opts for the first-order version – with the
Quinean quip in mind.
Another research programme that should not presuppose any mathematics is
logicism. In Frege’s flawed masterpiece, the Grundgesetze der Arithmetik3 [Basic
Laws of Arithmetic], he uses the logic he developed earlier in his Begriffsschrift4 to
derive the axioms of arithmetic. If the logic he uses for this is inclusive of
mathemat-ics, this logicist reduction of arithmetic to this logic would not be of the epistemic
merit that Frege envisioned for it. Logic is epistemically safe; to show that
arith-metic is so too, a reduction of the axioms of aritharith-metic to basic logical truths and
carried out by purely logical means would suffice. The system that Frege for the
first time in the history of logic introduces in theBegriffsschriftis polyadic predicate
logic, essentially the logic that we use today, modulo Frege’s unusual and perhaps
awkward seeming notation. It is also second-order.
Alas, Frege’s “logical system” of theGrundgesetze included one additional
“log-ical law” that is not present in the Begriffsschrift: the infamous Basic Law V.
Bertrand Russell showed just before the publication of Frege’s second volume of
the Grundgesetze, that Frege’s logic was not only inclusive of mathematics, but in
fact inclusive of everything. A contradication, known as Russell’s Paradox, is
deriv-able from Basic Law V. Frege briefly attempted to fix the problem, but gave up
quite quickly. The planned third volume of theGrundgesetze, for which the logicist
foundation for real, and perhaps complex analysis was planned, never appeared.
It has more recently been observed that Frege only uses Basic Law V to derive one
other principle, that already figures in his earlier philosophical outline of his logicist
reduction, the Grundlagen der Arithmetik5 [The Foundations of Arithmetic]:
so-called “Hume’s Principle”. Crispin Wright shows in his (Wright, 1983) that Frege’s
project can indeed be carried out, if Hume’s Principle is assumed. The
Peano-Dedekind axioms of arithmetic are derivable from it. Thus a neo-logicist research
programme of Neo-Fregeanism took wing.6 The project is sometimes also called
‘Abstractionism’; Hume’s Principle, and other principles of the same form are called
abstraction princples. These are conceived as implicit definitions that introduce on
their left-hand side of a bi-conditional a mathematical concept, like ‘natural number’
in the case of Hume’s Principle, and express on their right hand side an equivalence
relation in purely logical terms, a bijection between the extension of two predicates
in the case of Hume’s Principle. The logic that Neo-Fregeanism uses is second-order
logic, as it was for Frege. If second-order logic is indeed set theory, the project loses
almost all its interest: a reduction of arithmetic and analysis to set theory is no
news. In particular, it would not show that arithmetic inherits its epistemic status
from Hume’s Principle (which is arguably analytic). The ideal justification sought
for the axioms of arithmetic, and with them all arithmetical truths, requires that
the logic used is properlogic, and not set theory in disguise.
There are more projects in the philsophy of mathematics that also use
second-order logic. Geoffrey Hellman’s modal structuralism,7 for example, is formulated in
second-order logic with modal operators. Given the nominalist character that
Hell-man wants his approach to have, second-order logic had better not be mathematics
5(Frege, 1884).
6See (Wright, 1983), (Hale, 1987), (Hale and Wright, 2001); see also (MacBride, 2003) for an
excellent survey and critical discussion of the Neo-Fregean project.
if the project is to have any chance to succeed. Stewart Shapiro’s structuralism is
also dependent on second-order logic.8 As we will see, however, he does not believe
that a sharp line between mathematics and logic can be drawn, and that the case
of second-order logic especially shows this. My thesis, thus, not only has to argue
against the enemies of second-order logic, but against at least some of its proponents,
too.
1.2
Outline of the Thesis
After a short introduction to the formal system of second-order logic in chapter
2, my argument for the logicality of second-order logic begins, in chapter 3, with
investigating Quine’s reasons to claim that second-order logic is set theory. It will
turn out that Quine argues from ontology. The second-order quantifiers have to have
a range, and as they are quantifying into predicate position, they have to range over
some kind of universals. The best case for second-order logic, therefore, is according
to Quine that they range over sets, as these are the only universals (he sees sets as
such) that he can accept.
In chapter 4 I discuss George Boolos’ attempt to rebut the Quinean argument by
providing a plural interpretation of the order quantifiers. Monadic
second-order quantifiers can be interpreted as just ranging over the first-second-order domain, and
thus not introducing any new ontology, if they are construed as quantifying plurally.
Boolos provides a translation of the monadic second-order existential quantifier as
‘there are some things’, analogous to the first-order quantifier, ‘there is something’.
Boolos’ claim is that this way, no new ontological commitment arises.
Since according to my criticism of Boolos’ attempt to justify the second-order
quantifiers in this way – and of the attempt of his followers to patch up his account
where it is found wanting – leads me to reject this way of arguing against Quine,
I analyse in chapter 5 Quine’s criterion of ontological commitment in detail. The
criterion is found deficient, even for the paradigm cases of formalised first-order
theories that are the golden standard of all scientific and philosophical enterprise
for Quine. I suggest a natural modification and precisification of Quine’s criterion
to make it indeed applicable to all first-order theories. The resulting new criterion,
however, suggests that the introduction of second-order quantifiers does not bring
about any ontological commitment that was not already contained in the first-order
theory; in particular, no new commitment to sets arises.
The deductive system of second-order logic is incomplete with respect to its
standard semantics, in the sense that there are conclusions that are declared to
follow from some premises by the standard model-theoretic semantics, that are not
deducible from them in the deductive system, and indeed not inanysound deductive
system for this semantics. This is a corollary of G¨odel’s incompleteness theorem
for arithmetic. On the grounds of the lack of a complete proof procedure for the
semantical consequence relation, second-order logic is often denied the rank of a
proper logic. Surely, so the claim goes, every proper logic must be complete. I argue
in chapter 6 that there is no good reason to think so.
Chapter 7 deals with one of the main proponents of second-order logic today,
Stewart Shapiro. In his (Shapiro, 1991) he makes a case for second-order logic on the
basis of the claim that accepting second-order logic is the only way to make sense
of mathematical practice. His argument is, roughly, that mathematical practice
infinite structures, like the structure of the natural numbers or the structure of the
real numbers, and discern them. To do justice to this fact, Shapiro claims, one
has to consider categorical axiom systems for the mathematical theories that are
about these structures. A categorical axiom system is defined as one that has, up to
isomorphism, only one interpretation: all its models are isomorphic to each other. It
can be shown that no first-order axiomatisation of a theory over an infinite domain
can be categorical. The resources of second-order logic with its standard (but not a
Henkin) semantics allow us, however, to give categorical axiom systems of arithmetic
and real analysis, for example.
Critics of Shapiro argue that the standard model theory of second-order logic
makes strong mathematical presuppositions. While this presumably shows that
second-order logic can be no proper logic, if the model-theoretic system is identified
with second-order logic, this does not constitute an argument against Shapiro. He
rejects, precisely on these grounds the sharp distinction between logic and
math-ematics. His critics further argue that the categoricity results do not show the
determinacy that Shapiro requires for his argument, and some also suggest that
a second-order standard model-theoretic treatment hampers mathematical practice
rather than doing justice to it.
In chapter 8, finally, I introduce what I call the Deductivist Conception of Logic.
The model-theoretic approach to account for logical consequence is rejected on the
grounds that model theory is mathematical. As already argued above, however, a
proper logic must not presuppose any substantial, mathematical content since it
otherwise cannot fulfill his purpose to facilitate ideal justification. The Deductivist
proposal is to characterise logical consequence by purely deductive means. I argue
up until that point apply. Moreover, it appears that on a Deductivist conception
second-order quantification is sufficiently similar to and indeed of a piece, in a sense,
with first-order quantification that the former should count as properly logical if the
latter does.
My conclusion is, as already suggested in the first sentence of this introduction,
that second-order logic is proper logic, if it is construed in a Deductivist way. The
remainder of chapter 8 discusses further objections against second-order logic that
appear to be particularly pressing for the a Deductivist approach: an apparent
inherent incompleteness of second-order logic, that is not relative to some some
model-theoretic semantics, and the impredicativity of the second-order quantifiers.
An appendix discusses Etchemendy’s arguments concerning the concept of logical
consequence. Since the negative part of his project in which he criticises the Tarskian
reductive analysis of logical consequence is not unlike mine in spirit and character,
it seemed justified to include this appendix that discusses Etchemendy’s arguments
in some detail. Moreover, despite sharply critising the “interpretational”
model-theoretic semantics, Etchemendy come to a conclusion that is diametrically opposed
to the Deductivist account. He argues for what he calls “representational semantics”
which still uses (more or less) the standard model theory. A comparison of the two
Chapter 2
The Formal System of
Second-Order Logic
2.1
Preliminaries
This chapter introduces the formal system of classical second-order predicate logic.
Unless stated otherwise, the presentation follows chapters 3 and 4 of (Shapiro, 1991).
Standard systems of second-order logic can also be found in (Church, 1956) or
(Mendelson, 1997).1 A tree (or tableaux) system for second-order logic is introduced
in (Jeffrey, 1967); (Bell et al., 2001) also contains such a system.
2.2
Language
The language of a standard first-order logic is presupposed; see, for example,
(Mendel-son, 1997) or (Boolos and Jeffrey, 1985). Only the material conditional ‘⊃’, negation
1Note that only the 4th edition of Mendelson’s book contains a presentation of second-order
‘¬’, and the universal quantifier ‘∀’ are taken as primitive, the other logical constants
are defined in the usual way. The first-order existential quantifier, e.g., is defined as
∃xΦ =df ¬∀x¬Φ
where ‘Φ’ is a schematic letter standing for an arbitrary formula of the system.
For the language of second-order logic we introduce second-order variables: n
-place predicate variables, ‘Xn’, ‘Xn
1’, ‘X2n’, ‘Yn’, ‘Y1n’, ..., that can stand in place
of n-place predicate letters, and n-place function variables, ‘fn’, ‘fn
1,’‘f2n’, ..., that
can stand in the place of n-place function letters. The superscript indicates the
number n of argument places. Thus, ‘X41’ is a one-place predicate variable, ‘f13’
is a three-place function variable. In the following the superscripts indicating the
number of argument places will usually be omitted. Counting the terms that follow
the variable will disambiguate.
The language also contains second-order universal quantifiers that are formed
by attaching the ‘∀’ to second-order variables: ‘∀X’, ‘∀f’, etc. The second-order
existential quantifiers are defined as:
∃XΦ =df ¬∀X¬Φ
∃fΦ =df ¬∀f¬Φ
‘=’ is not taken as primitive but defined:
x=y =df ∀X(Xx≡Xy)
We can also define:
x6=y =df ¬x=y
The recursive formation rules that are added to those for the language of first-order
logic are:
If ‘f’ is ann-place function variable andphxinqis a sequence ofn terms, then
pfhxinq is a term.
If ‘R’ is an n-place predicate variable and phxinq is a sequence of n terms,
then pRhxinq is an atomic formula.
If ‘f’ is a function variable and Φ is a formula, then p∀f(Φ)q is a formula.
If ‘R’ is a predicate variable and Φ is a formula, then p∀R(Φ)q is a formula.
Convention: The parenthesis that inclose the formula and indicate the scope
of the quantifier can be omitted in cases where there is no scope-ambiguity.
2.3
Deductive Systems
2.3.1
Axiomatic System
Let us define a standardaxiomatic system for second-order logic first.
Let Γ be a set of sentences of the language and Φ a single sentence of the language.
Define a deduction of Φ from Γ to be a finite sequence Φ1, ...,Φn such that Φn is Φ
and, for eachi6n, Φi is an axiom (see below), or Φifollows from previous sentences
in the sequence by one of the rules of inference (see below). We can symbolise that
there is a deduction of Φ from Γ as: Γ`Φ.
Let a proof of Φ be a deduction of Φ from the empty set. Call Φ a theorem if
The following are axiom schemata. Any formula obtained by substituting
for-mulas for the schematic letters ‘Φ’, ‘Ψ’, and ‘Ξ’, is anaxiom of the system.
Φ⊃(Ψ ⊃Φ)
(Φ⊃(Ψ⊃Ξ)) ⊃((Φ⊃Ψ)⊃(Φ⊃Ξ))
(¬Φ⊃ ¬Ψ)⊃(Ψ ⊃Φ)
∀xΦ(x)⊃Φ(t)
where ‘t’ is a term free for ‘x’ in Φ
∀XnΦ(Xn)⊃Φ(T)
where ‘T’ is an n-place predicate letter free for ‘Xn’ in Φ
∀fnΦ(fn)⊃Φ(p)
where ‘p’ is ann-place function letter free for ‘fn’ in Φ
A term ‘t’ is free for ‘x’ in Φ if no variable has an occurrence that is both free in ‘t’
and bound inpΦ(t)q; analogously for predicate and function letters.
The rules of inference of the system are:
Modus ponens:
from Φ and pΦ⊃Ψq infer Ψ
Generalisation:
from pΦ⊃Ψ(t)q inferpΦ⊃ ∀xΨ(x)q
provided ‘t’ does not occur free in Φ or in any of the premises of the deduction
from pΦ⊃Ψ(T)qinfer pΦ⊃ ∀XΨ(X)q
from pΦ⊃Ψ(p)qinfer pΦ⊃ ∀fΨ(f)q
provided ‘f’ does not occur free in Φ or in any of the premises of the deduction
The usual axioms and rules for the other sentential connectives and the existential
quantifier, which are defined here and not taken primitive, can be derived from the
rules and axioms given above.
We add to the system an axiom schema of comprehension:
∃Xn∀hxin(Xnhxin ≡Φhxin)
provided ‘Xn’ does not occur free in Φ;phxinqis a sequence ofnfirst-order variables;
p∀hxinq abbreviates a sequence ofn quantifiers p∀xiq, for 16i6n.
The comprehension schema asserts that every open sentence of the language there
a (possibly many-place) predicate with the same extension. If Φ contains no bound
second-order variables, we call the corresponding instance of the comprehension
schema predicative, and impredicative otherwise.
We also add an axiom of comprehension for functions:
∃Xn+1(∀hxi
n∃!yXn+1hxiny ⊃ ∃fn∀hxinXn+1hxinfhxin)
Shapiro suggests to add instead of the comprehension for functions a stronger
prin-ciple, an axiom of choice. He writes:
The axiom of choice has a long and troubled history [...], but it is now
essential to most branches of mathematics. In fact, a corresponding
meta-theoretic principle is necessary for many of the theorems reported
[in (Shapiro, 1991)]. Mathematical logic also thrives on the axiom of
choice.2
The axiom of choice is:
∃Xn+1(∀hxin∃yXn+1hxiny ⊃ ∃fn∀hxinXn+1hxinfhxin)
Note, that it does not have the uniqueness condition attached to the first existential
quantifier that the axiom of comprehension for functions has. The antecedent of
the axiom of choice asserts that for each sequence phxinq there is at least (exactly,
for the weaker comprehension for function above) one ‘y’ such that the sequence
phxinyq satisfies pXn+1q. The consequent asserts the existence of a function that
“picks out” one such ‘y’ for each phxinq.
The axiom of choice, which is often considered problematic, cannot be discussed
here.
2.3.2
Natural Deduction
Equivalently, we can give a natural deduction system for second-order logic. (Shapiro,
1991) does not contain such a system; for a system similar to the one introduced
here, see (Prawitz, 1965).
The classical introduction- and elimination-rules for the propositional fragment
are presupposed (see for example (Prawitz, 1965)).
Observe, that function letters are dispensable. This is also the case for the
axiomatic system, of course, but as I will, for simplicity’s sake, use quantification
into function-letter-position in some of the chapters below, it seemed advisable to
introduce them into the language and axiomatic system. (n+ 1)-place predicate
variables can serve as surrogates forn-place function variables, however. The clause
p∀hxin∃!yFn+1hxinyq indicates that pFn+1qis in effect an n-place function.
and quantifiers. The introduction (I) and elimination (E) rules for the first- (∀1)
and second-order (∀2) universal quantifiers are:
Φ(t)
∀xΦ(x)∀
1-I Φ(T)
∀XnΦ(Xn)∀ 2-I
∀xΦ(x)
Φ(t) ∀
1-E ∀X
n
Φ(Xn)
Φ(Ξ) ∀
2-E
Φ is an open sentence matching the number of argument places of the expression it
applies to, in the case of the second-order rules possibly just one term; ‘t’ is a term
of the language.
Restrictions:
∀1-I ‘t’ does not occur free in any of the assumptions that
pΦ(t)qdepends on.
∀1-E ‘t’ is free for ‘x’ in Φ.
∀2-I ‘T’ is a n-place predicate letter and does not free in any of the assumptions
that pΦ(T)qdepends on.
∀2-E Ξ is an open sentence with n argument places; no variable in Ξ is bound in
2.4
Semantics
2.4.1
Semantics for First-Order Logic
The model-theoretic semantics for the first-order logic fragment of second-order logic
is presupposed, and merely sketched here.
A model is an order pair M = hd, Ii, in which d is the domain of the model,
a non-empty set, and I is an interpretation function that assigns objects in d and
sets that are constructed from objects in d to the non-logical vocabulary of the
language. If ‘a’ is a term of the language, for example, I(‘a’) is a member of d; if
‘R’ is a two-place predicate letter, I(‘R’) is a subset of d×d (‘×’ standing for the
Cartesian product). Avariable assignment s is a function from the variables of the
language tod.
For each model and variable assignment there is a denotation function that
as-signs an object in d to every term of the language. Satisfaction is defined in the
usual way as a relation that holds between models, variable assignments, and
formu-lae. Let us write ‘M, s Φ’ for ‘M and s satisfy Φ’. If M, s Φ for every variable
assignments, we say thatM is amodelof Φ. Ifsands0are two variable assignments that agree on all free variables of Φ, thenM, sΦ if, and only if, M, s0 Φ. Since if Φ is a sentences, i.e. a formula with no free variables, the variable assignment
makes no difference, we can just write M Φ.
A formula Φ is satisfiable if, and only if, there is a model M and a variable
assignment s on M such that M, s Φ. A set of formulae Γ is satisfiable if, and
only if, there is a model M and a variable assignment s on M such that M, s Φ
for every Φ∈Γ. A formula Φ is asemantic consequenceof a set of formulae Γ, if the
can also say that Φ is a semantic consequence of Γ if, and only if, for every model
M and any variable assignment s on M, ifM, s Ψ for any Ψ∈Γ, then M, sΦ.
We can also write this as ‘ΓΦ’.
A formula Φ is validif, and only if, M, sΦ for all M and alls onM. We can
also write this as ‘Φ’.
2.4.2
Standard Semantics for Second-Order Logic
We can build a standard semantics for second-order logic on this basis. A standard
model is still an ordered pair hd, Ii, as in first-order logic. A variable assignment
is a function that assigns a member of d to each first-order variable, a subset of
dn to every n-place predicate variable, and a function from dn tod to each n-place
function variable. (dn is d for n = 1, d×d for n = 2, d×d×d for n = 3, etc.)
The range of the one-place predicate letters is thus the powerset of the domain d;
generally the powerset ofdn is the range of the n-place predicate letters.
The new clause for the denotation function which is to be added to those for
first-order logic is:
Let M =hd, Ii be a model and s be a variable assignment onM. The
deno-tation of fnhti
n in M, s is the value of the function s(fn) at the sequence of
members of d denoted byhtin.
The relation of satisfaction is also extended from first-order logic. The three new
clauses are:
If Xn is an n-place predicate variable andhti
n is a sequence of n terms, then
M, sXnhti
n if, and only if, the sequence of members ofd denoted byhtin is
M, s∀XΦ if, and only if, M, s0 ∀XΦ for every s0 on M that agrees with s
at every variable except maximally X.
M, s ∀fΦ if, and only if, M, s0 ∀XΦ for every s0 onM that agrees with s
at every variable except maximally f.
The definitions of satisfiability, semantic consequence and validity remain the way
they are introduced for first-order logic.
2.4.3
Henkin Semantics
In a Henkin semantics (introduced by Leon Henkin in (Henkin, 1950)) it is not
assumed that the n-place predicate variables range of the full powerset of dn, but
a separate domain is specified for them in each model, and also for the function
variables. A Henkin model is a quadruple MH = hd, D, F, Ii in which d is the
domain and I an interpretation function as above. D is a sequence of non-empty
sets D(n) that contain subsets of dn for every n, to be assigned to the n-place
predicate variables as we will see below. Likewise, F is a sequence of non-empty
setsF(n) of functions from dntod. Intuitively, the range of the one-place predicate
variables, for example, is a fixed subset of the powerset ofd for each model.
Avariable assignmentis a function that assigns a member ofdto each first-order
variable, a member ofD(n) to eachn-place predicate variable, and a member ofF(n)
to eachn-place function variable. The remaining features of a Henkin semantics are
analogous to those of the standard semantics.
The four new clauses, to be added to the semantics of first-order logic, are:
LetMH =hd, D, F, Iibe a Henkin model andsa variable assignment onMH.
The denotation of fnhti
sequence of members of d denoted by htin.
If Xn is an n-place predicate variable andhtin is a sequence of n terms, then
MH, s Xnhtin if, and only if, the sequence of members of d denoted by htin
is a member of s(Xn).
MH, s ∀XΦ if, and only if, MH, s0
∀XΦ for every s0 on MH that agrees
with s at every variable except maximallyX.
MH, s ∀fΦ if, and only if, MH, s0
∀XΦ for every s0 on MH that agrees
with s at every variable except maximallyf.
Again, the definitions of satisfiability, semantic consequence and validity are
anal-ogous to those introduced for first-order logic, only that they are with respect to
Henkin models.
2.5
Some Meta-Theoretic Results
In this section the meta-theoretic results concerning first- and second-order logic
that will be of interest for the philosophical discussion in the following chapters are
stated. The results are discussed where they are mentioned in the later chapters.
Their proofs are omitted. For first-order logic, they can be found, for example, in
(Mendelson, 1997) or (Boolos and Jeffrey, 1985); for second-order logic, the proofs
are sketched in (Shapiro, 1991).
2.5.1
Standard and Henkin Semantics
A Henkin model in which all the D(n) are the full powerset of dn, and all the F(n)
the sets of all n-place functions from dn to d is obviously equivalent to a standard
model. Thus, if we restrict the range of Henkin model to such models, this restricted
Henkin semantics will be equivalent to the standard semantics. It hence follows that:
If Φ is valid according to Henkin semantics, then Φ is valid according to the
standard semantics.
If Φ is a semantic consequence of Γ according to Henkin semantics, then Φ is
a semantic consequence of Γ according to the standard semantics.
If Φ is satisfiable according to the standard semantics, then Φ is satisfiable
according to Henkin semantics.
The converse does not hold in any of the cases.
2.5.2
First-Order Logic
The soundness and completeness theorems for first-order logic are well know. They
are:
Soundness: Let Γ be a set of formulae and Φ a formula of the first-order language.
If Φ `Γ then ΦΓ. A fortiori, if `Γ then Γ.
Completeness: Let Γ be a set of formulae and Φ a formula of the first-order
language. If Φ Γ then ΦΓ. A fortiori, if Γ then `Γ.
is straightforward. One checks each axiom and rule of inference.
Vir-tually no substantial set-theoretical assumptions are needed. [...] [T]he
completeness of first-order logic depends on a principle of infinity (in the
metalanguage). If the model-theoretic semantics had no models with
infinite domains, the completeness theorem would be false.3
Another standard meta-theoretical results is compactness:
Compactness: Let Γ be a set of formulae of the first-order language. If every finite
subset of Γ is satisfiable, then Γ is satisfiable.
It follows that if an infinite set of first-order formulae is non satisfiable, it has a
finite subset that is not satisfiable. The compactness theorem is a direct corollary
of soundness and completeness. For the remaining two standard meta-theorems we
should introduce more technical terminology. Let M = hd, Ii and M0 = hd0, I0i
be two models. We define that M0 to be a submodel of M if, and only if, d0 is a subset ofd,I and I0 give the same denotation to each individual constant, and the interpretation of each predicate and function letter underI0 is the restriction tod0 of the corresponding interpretation underI. If the theory contains function constants,
then d0 must be closed under these functions.
L¨owenheim-Skolem theorem: If M is a model of a set Γ of first-order formulae,
then M has a submodel M0 whose domain is at most countable infinite, such that for each assignment s on M0 and each formula Φ in Γ: M, s Φ if, and only if, M0, sΦ.
The axiom of choice is required in the meta-theory for the proof of the L¨
owenheim-Skolem theorem.
L¨owenheim-Skolem-Tarski theorem: Let Γ be a set of first-order formulae. If,
for each n ∈ ω, there is a model of Γ whose domain has at least n members,
then for any infinite cardinal κ, there is a model of Γ whose domain has at
least cardinality κ.
This entails that every first-order theory with a countably infinite model, e.g. Peano
Arithmetic, has an uncountable model, too. By the L¨owenheim-Skolem theorem,
real analysis which has as the intended uncountable domain the real numbers, has
a countable model.
2.5.3
Second-Order Logic with Standard Semantics
One, and only one, of these results for first-order logic carry over to second-order
logic with standard semantics:
Soundess: Let Γ be a set of formulae and Φ a formula of the second-order language.
If Φ ` Γ then Φ Γ according to the standard semantics. A fortiori, if ` Γ
then Γ according to the standard semantics.
The proof involves the assumption that every formula in the meta-theory determines
a set, and uses the principle of separation (Aussonderung). If we add the axiom of
choice to the deductive system (as Shapiro suggests), then we need the axiom of
choice in the meta-theory, too.
As mentioned above, first-order theories with countable models also have
un-countable models, and first-order theories with unun-countable models also have
count-able models. A way to paraphrase this is that first-order theories cannot determine
that their domain has a certain infinite cardinality. Second-order theories,
axiomatic system is categoricalif, and only if, all of its models are isomorphic. The
theory of second-order arithmetic, for example, is categorical given these axioms:
∀x(sx6= 0) (zero)
∀x∀y(sx=sy⊃x=y) (successor)
∀X[(X0∧ ∀x(Xx⊃Xsx)⊃ ∀xXx] (induction)
‘0’ and ‘s’ are non-logical constants for zero and the successor function, respectively.
Addition and multiplication do not have to be mentioned in the axioms, as they
can be defined in the second-order theory. LetAR be the conjunction of the three
axioms above.
Categoricity of second-order arithmetic: LetM1 =hd1, I1iandM2 =hd2, I2i
be two models of second-order arithmetic (with the axioms mentioned above).
For 1 6 i 6 2 let 0i be the interpretation of zero in di, and let si be the
interpretation of successor. If M1AR andM2AR, thenM1 andM2 are
isomorphic: there is a bijection f, a one-to-one function fromd1 ontod2, such
that f(01) = 02, and for each a∈d1,f(s1(a)) =s2(f(a)). That is,f preserves
the structure of the models.
Since the intended interpretation, the natural numbers, is countably infinite and a
model ofAR, it follows from categoricity that all models of second-order arithmetic
are countably infinite. The analogous result holds for real analysis in its
second-order axiomatisation. All of its model are of the cardinality of the continuum, i.e. of
the powerset of the natural numbers.
For second-order Zermelo-Fraenkel set theory a similar, but restricted result
holds, often calledquasi-categoricity. Intuitively, the quasi-categoricity of this theory
or one is isomorphic to an “initial segment” of the other. All models of
second-order Zermelo-Fraenkel set theory are isomorphic up to an inaccessible rank (the
existence of inaccessible cardinals is independent of this theory). Thus, all models,
in a sense, “agree on” the structure below the least inaccessible rank, or, as it is
sometimes glossed, any two models are isomorphic up to the least inaccessible rank.
The theory that we get from adding the claim that there arenoinaccessible cardinals
to second-order Zermelo-Fraenkel set theory is categorical.
It follows immediately from categoricity that both theL¨owenheim-Skolemand
L¨owenheim-Skolem-Tarski theorems fail for second-order logic with standard
semantics. Also Compactness failsas is shown to follow from the categoricity of
second-order arithmetic in my chapter 7 below.
Moreover, it follows from G¨odels incompleteness theorem for arithmetic that
there is no sound deductive system that is complete with respect to the standard
semantics. Thus, second-order logic with standard semantics is inherently
incom-plete. This is easy to see: Take the G¨odel sentence of the second-order axiom
system of arithmetic, call it G. By G¨odel’s proof, G is true in the theory, but not
provable in the deductive system. pAR ⊃ Gq, then, is not provable either in the
deductive system, but it is a validity of the standard semantics of second-order logic
as follows from the categoricity result for arithmetic. This, however, holds for any
deductive system.
2.5.4
Second-Order Logic with Henkin Semantics
Second-order logic with Henkin semantics has the same meta-logical properties as
the-ory that has an infinite domain, because of the L¨owenheim-Skolem, and L¨
owenheim-Skolem-Tarski theorems.
The range of Henkin models has to be restricted (in a straightforward way) in
order to be able to prove that the meta-theorems hold. Define a Henkin model to be
faithfulto the deductive system of second-order logic if, and only if, it satisfies every
instance of the comprehension schema (and the axiom of choice, if this is added).
Soundness: If Γ ` Φ, then Φ is satisfied by every faithful Henkin model that
satisfies every member of Γ. A fortiori, if ` Φ, then Φ is satisfied by every
faithful Henkin model.
Completeness: Let Γ be a set of formulae and Φ a formula. If M, sΦ for every
faithful Henkin model M and variable assignment s onM that satisfies every
member of Γ, then Γ`Φ.
Compactness: Let Γ be a set of formulae. If every finite subset of Γ is satisfiable
in a faithful Henkin model, then Γ is satisfiable in a faithful Henkin model.
For the L¨owenheim-Skolem theorem the notion of a submodel has to be extended.
We also have to define acorrespondence functionbetween a submodel and a model.
Intuitively, this function maps the sets that are assigned to the predicates and
functions in the submodel to those in the model. Details are omitted here; they can
be found in (Shapiro, 1991), pp. 92–94. It suffices to say here that the analogues
for theL¨owenheim-Skolem andL¨owenheim-Skolem-Tarski theoremsholdfor
Chapter 3
On Quine
3.1
Introduction
An intuitive way to think of second-order logic is to add upper case variables figuring
in predicate position to the standard first-order logic and allow for the binding of
these with the usual existential and universal quantifier. The inferential behaviour
of the second-order quantifiers might be taken to be sufficiently analogous to that
of the first-order quantifiers. Let us further take for granted that first-order logic
is logic proper. Thus, if first-order logic is proper logic, and if what is added to it
to get second-order logic is not fundamentally different from what we had before, it
appears that second-order logic is proper logic, too.
The most famous critic of the claim that second-order logic is proper logic is
probably W.V. Quine. He attacked second-order logic vigorously over decades on
various grounds. In his writings he ascribed to it an air of incoherence, found it
sheep’s clothing”.1 Obviously, Quine cannot hold all these claims together. I will in
this chapter reconstruct how rather one of these claims leads to the next, and present
Quinean reasons as to why. In presenting different conceptions of what is happening
when one quantifies into predicate position, I will argue, Quine ends up with stating
that the best case that can be made for second-order logic is taking it to be some
sort of class theory (or set theory: Quine uses ‘class’ and ‘set’ interchangeably2).
This, however, attracts a charge akin to intellectual dishonesty. Being a class theory,
second-order logic has an ontological commitment to classes, but this commitment
is masked in the form of second-order quantification; the ontological commitment is
not made explicit. Second-order logic, for Quine, is hence practically a Trojan Horse,
and a gigantic one: in the first edition of hisPhilosophy of Logic Quine ascribes to
second-order logic a commitment to the “staggering existential assumptions” of set
theory.3
These claims, how they work together, and how a Quinean argument can be
rationally reconstructed is the topic of this chapter. Quine’s criterion for ontological
commitment plays a major in various parts of the argument. For the purpose of this
chapter, this criterion is granted. Chapters 4 and 5, though, discuss this assumption
in detail, and the criterion is finally rejected in chapter 5.
Before going into the exposition and discussion of Quine’s quarrels, it might be
worth mentioning a possible psychological explanation as to why Quine believes that
second-order logic is set theory. Quine’s way to think about second-order logic is
1So the title of a section in (Quine, 1986a), pp. 64–66.
2Quine takes the use of the term ‘set’ rather than ‘class’ in mathematical circles to be almost
entirely a matter of mere fashion. This does not mean, however, that he neglects the distinction
between so-calledproper classes(Quine prefers the term ‘ultimate class’) and such classes that can
themselves be members of other classes. See (Quine, 1969b), p. 3.
to conceive of it as a variant of the Simple Theory of Types. A quote from his Set
Theory and its Logic makes this clear:
[An] assimilation of set theory to logic is seen also in the terminology
used by Hilbert and Ackermann and their followers for the fragmentary
theories in which the types leave off after finitely many. Such a theory
came to be called the predicate calculus (Church: functional calculus)
of nth order [...], where n is how high the types go. Thus the theory
of individuals and classes of individuals and relations of individuals was
called the second-order predicate calculus, and seen simply as
quantifica-tion theory with predicate letters admitted to quantifiers. Quantificaquantifica-tion
theory proper came to be called the first-order predicate calculus.
This is a regrettable trend. Along with obscuring the important
cleav-age between logic and “the theory of types” (meaning set theory with
types), it fostered an exaggerated if foggy notion of the difference
be-tween the theory of types and “set theory” (meaning set theory without
types) – as if the one did not involve outright assumptions of sets the
way the other does.4
Also, as further evidence for my suggestion, in Quine’s earlier papers where he
raises his complaints about higher-order quantification, he indeed explicitly mentions
Russell’s Simple Theory of Types, e.g. in (Quine, 1947), rather than second-order
predicate logic which is his target especially in hisPhilosophy of Logic(Quine, 1970).
The criticism concerning one or the other shows continuity; only the terminology
changes.
While there are interesting structural similarities between the Simple Theory
of Types and higher-order logic, which can often usefully be exploited, it is still
important to keep them apart. The Theory of Types Quine is concerned with is
explicitly designed to be a theory of classes5, with typed class variables, while
higher-order logic allows quantification into predicate position. In particular, variables for
many-place predicates (or relation symbols) and their binding with quantifiers do
not figure in this type theoretical system. The latter are, however, an important
ingredient in higher-order logic. Neglecting this can lead into trouble: many-place
predicates will play an interesting and surprising role further down in this chapter,
and also in chapters 4 and 5.
Running a Quinean conception of the Theory of Types and higher-order logic
together would provide an excellent ground to claim that second-order logic is really
a theory of classes, were it not fallacious to do so. The philosophical arguments for
the claim that second-order logic is ontologically committed to sets are, of course,
independent of these anecdotical remarks about Quine. Bearing in mind that this
is Quine’s viewpoint, however, sometimes helps understanding why Quine expresses
things the way he does, especially in the earlier papers, which sometimes seems a bit
awkward to the reader today who is more familiar with a conception of second-order
logic that is quite independent of the Theory of Types.
3.2
Incompleteness and Branching Quantifiers
Before discussing the Quinean worries mentioned above, it is worth noting an
ob-jection that is almost unrelated to Quine’s other complaints against second-order
logic: the common objection concerning its incompleteness. There are some places
in his Philosophy of Logic where Quine suggests that the lack of a completeness
proof indicates that the border to set theory, i.e. mathematics, has been crossed.
Most notably this comes up in his discussion of branching quantifiers.6 It appears,
however, that the only place in Quine’s writings where he mentions this worry in
connection with second-order logic is in one of his replies in his volume of Schilpp’s
Library of Living Philosophers7 – and even there the incompleteness of second-order
logic is only mentioned in one sentence, and not discussed further. (My chapter 6
below contains a detailed discussion of the incompleteness objection.)
It might be that Quine takes it that what he offers against second-order logic is
devastating enough so that there is no need for the additional objection. Another
explanation would be that he does not think that the argument from
incomplete-ness is a particularly strong one. Trying to reconstruct Quine’s view on this, one
notices that he is not particularly fond of the model theoretic approach to logic,
as becomes clear from his discussion of it in chapter 4 of his Philosophy of Logic.8
Quine’s preferred way to characterise logical truth is substitutional. He defines a
logical truth as a substitution instance of a valid logical schema, where a valid
logi-cal schema is one that has only true substitution instances.9 Logical schemata, for
Quine, are sentence-like well-formed formulae, constructed according to the
syntac-tical rules of first-order logic. His definition of logical truth obviously is hostage to
the identification of the logical vocabulary about which Quine says precious little.
Quine’s worry about the model-theoretic approach to logical truth is that it
6(Quine, 1986a), p. 90–91.
7(Quine, 1986b), p. 646.
8(Quine, 1986a), pp. 51–53.
requires a set-theoretical interpretation of the language,10 and thus affords an
on-tological commitment to sets that his substitutional account apparently avoids. On
grounds of ontological parsimony Quine’s substitutional account is preferable to the
model-theoretic one.11 Quine concedes, however, that he does not take his
substitu-tional account to bewholly independent of sets. It deals with sentences, and Quine
takes sentences to be sets of their tokens. He also claims that sets are necessary
for the construction of a syntax for a language, especially in order to be able to
talk about arbitrarily long sentences even if some will never be written down.12
Al-though Quine is prepared to commit himself to sets in this way, he takes it that the
commitment is rather modest compared to the one that the model-theoretic account
brings with it. The substitutional account, Quine claims, merely requires finite set
theory:
The way to look upon the retreat [from model theory to the
substitu-tional account], then, is this: it renders the notions of validity and logical
truth independent of all but a modest bit of set theory; independent of
the higher flights.13
Quine does not say this explicitly, but one might take the underlying reasoning to
10(Quine, 1986a), p. 51. That I am sympathetic to Quine’s worries here, albeit for different
reasons, will become clear in chapters 7 and 8.
11(Quine, 1986a), pp. 55–56.
12For all of these concessions Quine makes concerning the commitment to sets, however, there
seems to be logical space for resistance. Syntax and proof theory are to some, perhaps sufficient, extent available without recourse to sets, as Quine himself argued together with Nelson Goodman in an early joint paper: (Goodman and Quine, 1947). Also other strategies have been proposed.
13(Quine, 1986a), p. 56. Quine’s exposition of this claim makes the detour via the possibility of
a G¨odel coding and an arithmetisation of syntax. Quine does not give the details, and they need
not concern us here either. In any case, he takes number theory “in effect equivalent still to a certain amount of set theory [...], [but] it is a modest part: the theory of finite sets.” Quine takes set theory to be all one needs as a foundation of mathematics, as numbers, functions and relations are definable as certain sets of sets. See, for example, (Quine, 1947), p. 79 (reappears as (Quine,
be that whatever formal language one uses, a certain modest amount of set theory
will be needed. Model-theoretic approaches, however, need in addition substantially
more set theory. Quine’s substitutional account does not need this additional bit
and is therefore to be preferred.
In section 3.4 below I will reconstruct and discuss Quine’s argument for the claim
that second-order logic is committed to the “staggering ontology” of set theory. If
the interpretation I give of Quine’s argument is right, set theory is not available in
bits as far as Quinean ontological commitment is concerned. If Quine, therefore,
wanted to uphold the claim that second-order logic is committed to the whole of
the set theoretical hierarchy, then he cannot uphold his claim that the
substitu-tional account fares better than the model-theoretic one on counts of parsimony:
commitment to some sets always means commitment to the entire hierarchy.
Be that as it may, the standard completeness proof shows that a given syntactic
system captures all the model-theoretic validities, and declares them to be theorems
of the system – the logical truths in the case of a system of logic.14 If, however, model
theory is not assigned a special role, the significance of such a proof is doubtful.15
Not only does Quine not assign any particularly special role to model theory, he is
also rather wary of it because of its use of set theory, as mentioned above. The lack
of a completeness proof for second-order logic is therefore something that Quine is
in no position to put forward as a strong argument.
Indeed, also in his discussion of branching quantifiers the incompleteness
ob-jection appears as a kind of add-on to his main criticism. Quine observes that a
14At least this is the part of it that Quine focuses on. The standard completeness proof shows not
only that the valid sentences are captured, but shows this generally for the consequences relation. Logical truths are merely a special case of this: consequences of an empty set of premises.
sentence like:
∀x∃y
∀z∃w
F(x, y, z, w)
is not equivalent to any of its first-order linear versions like:
∀x∃y∀z∃w F(x, y, z, w)
or:
∀z∃w∀x∃y F(x, y, z, w)
Rather, we need to quantify over functions, to get it “back into line”:16
∃f∃g∀x∀y F(x, f(x), y, g(y))
Since we are thus committed to functions (as we quantify over them), we have crossed
the border to mathematics, Quine claims: “We leave logic and ascend into
math-ematics of functions, which can be reduced to set theory but not to pure logic.”17
Quantification over functions, whoever, is second-order quantification. Quine uses
here second-order quantification to discredit (on ontological grounds) an alternative
logic, while his line on second-order logic is that it is dishonest at best, and otherwise
unintelligible, or even inconsistent.
This puts Quine into a strange position: of the two arguments he presents against
branching quantifiers, one relies on second-order logic which he emphatically rejects,
and the other one is from incompleteness which at least he cannot coherently consider
to be a very strong argument. To be charitable, Quine would probably retreat to a
rendering of the offendingly untidy branching formulae in linear first-order set theory.
The fact remains, however, that he does not do it that way, but rather recasts the
branching quantifiers formulae in second-order logic to make his point. Presumably,
he does so for didactical reasons: the connection between the branching, and the
second-order sentence can readily be seen, while a rendering in set theoretical terms
would be much more cumbersome. (Functions are defined as special sets of ordered
pairs; and an ordered pairhx, yi is defined as a set {{x},{x, y}}, following Wiener
and Kuratowski.18 If we have to use these sets and also spell out the conditions for
functionality the result will be much more difficult to parse,19 and its connection
to the branching sentence will not be as obvious.) This didactic strategy, however,
will not work if the utilised means is inconsistent or unintelligible. If Quine is right
about second-order logic, it seems as if he, hence, would have to accuse himself of
dishonesty. This certainly does not constitute a decisive counter-argument against
Quine’s position; the curious situation Quine got himself into nevertheless seems
noteworthy.
As already mentioned, the incompleteness allegation against second-order logic
is discussed in detail in chapter 6. In the rest of this chapter I attempt to unpack
Quine’s animadversions against second-order logic.
18Quine regards the Wiener-Kuratowski definition of an ordered pair as a paradigm case of a
successful explication: (Quine, 1960),§53; see also (Quine, 1947), p. 79.
19In a different context, the Wiener-Kuratowski definition of an ordered pair is spelled out in
3.3
Incoherence and Unintelligibility
3.3.1
Russell’s Paradox
In some of his earlier publications Quine alludes to Russell’s Paradox in his discussion
of higher-order quantification. Russell’s Paradox arises in na¨ıve set theory in the
following way. Let us say that for any open sentence there is a set that contains all
and only the objects that satisfy this open sentence. Prima faciethis sounds like a
reasonable suggestion, but this principle leads to Russell’s famous antinomy. Take
the predicate ‘being a set that does not contain itself’. Call the set that contains all
and only sets that do not contain themselves ‘r’ – the “Russell Set”. Doesrcontain
itself? It seems it cannot, sincer contains only sets that do not contain themselves;
for if it contained itself, it could not be amongst those. So it does not contain
itself. In that case, however, it is one of those sets that do not contain themselves,
and hence it has to be in the respective set, i.e. in itself. So, if r contains itself,
it does not, and if it does not contain itself, it does; or formally, r ∈ r ≡ ¬r ∈ r:
contradiction.
In On Universals (Quine, 1947) Quine suggests that an antinomy similar to
Russell’s for na¨ıve set theory might occur if one takes second-order quantification
seriously. If we allow binding predicate letters with quantifiers this means that they
“acquire the status of variables”, and that we are thus “granting [them] all privileges
of ‘x’, ‘y’, etc.”, i.e. of the first-order variables. This means allowing second-order
variables to occur in name position, and hence allowing formulae like ‘GH’ (just as
‘Gx’) “seems a very natural way of proclaiming a realm of universals” to Quine.20
He proceeds by giving a proof in this so characterised system of an analogue of
Russell’s Paradox:21
(1) GH ≡GH logical truth
(2) ∀H(GH ≡GH) (1), universal generalisation
(3) ∀F¬∀H(F H ≡GH)⊃ ¬∀H(GH ≡GH) instance of the ‘∀’-axiom22
(4) ¬∀F¬∀H(F H ≡GH) (2), (3), modus tollens
(5) ∃F∀H(F H ≡GH) (4), quantifier conversion
(6) ∃F∀H(F H ≡ ¬HH) (5), subst. ‘¬HH’ for ‘GH’
(7) ∃F(F F ≡ ¬F F) (6), “by a few easy steps”
Quine makes the detour through line (3) and (4) as the deductive system he proposes
in this paper does not have any inference rules for the existential quantifier; he treats
‘∃x’ as a mere abbreviation of ‘¬∀x¬’. Otherwise line (5) would follow immediately
from line (2) by existential generalisation – if we ignore for the moment that none of
the formulae is well-formed according to any standard higher-order logic or theory
of types. (I will come back to the question of peculiar syntax below).
This, however, is not the only oddity about this derivation. The substitution step
from line (5) to line (6) is invalid: it brings the first ‘H’ in the substituted expression
‘¬HH’ into the scope of the universal quantifier. Quine justifies this step by referring
to two of the rules of inference that he introduced earlier. These are the already
mentioned rule to allow the predicate letters all privileges of individual variables, and
the rule for substitution of formulae: “Substitute any formulae for ‘p’, ‘q’, ‘F x’, ‘F y’,
21Compare (Quine, 1947), p. 78. The notation is changed to match the symbolisation used
throughout this thesis. Moreover, Quine’s commentary column is omitted in favour of my own annotations, as Quine’s abbreviated comments would require rather extensive introduction. Only the annotation in line (7) is verbatim.